Detection-aware disparity refinement for weak-texture pavement target reconstruction in stereo vision

Abstract Accurate detection and depth perception of pavement targets from close-range remote sensing observations are important for intelligent road monitoring and transportation infrastructure assessment. However, existing stereo matching methods often produce unstable disparity estimates in weak-texture and low-height pavement regions, such as manhole covers and speed bumps, where the disparity distribution is easily affected by matching ambiguity and image noise. To address this issue, a detection-guided disparity post-processing strategy named Detection-Aware Disparity Refinement (DADR) is proposed. A custom pavement dataset containing approximately 3000 annotated images is first constructed from a binocular vision-based remote sensing platform, covering five pavement target categories, including manhole covers, speed bumps, road damage, storm drains, and iron plates. YOLOv11 is employed to detect target regions, while FoundationStereo is used to generate initial disparity maps from calibrated stereo images. Instead of modifying the stereo matching network, the proposed method performs a rule-based post-processing operation that combines detection bounding boxes, a local ground-reference disparity, and confidence-weighted fusion to stabilize the disparity representation within detected target regions. Experimental results show that the YOLOv11 detector achieves a Precision of 0.89, a Recall of 0.88, and an mAP@50 of 0.895, and compares favourably with other one-stage detectors trained under the same protocol. Visual analysis with magnified regions of interest shows that the proposed strategy produces more spatially coherent disparity responses for weak-texture pavement targets, and an ablation study identifies the local ground reference and the adaptive confidence weighting as the main contributing components. Quantitative evaluation indicates that the within-region disparity standard deviation decreases from 3.81 px to 1.25 px after refinement. Because no independent depth ground truth is available, this result is reported as improved local disparity consistency rather than improved absolute depth accuracy. The proposed approach offers a lightweight, detector-agnostic disparity stabilization strategy for low-height pavement targets in close-range remote sensing scenarios and provides useful geometric cues for subsequent road-condition assessment, while its reliance on a locally planar ground reference and on detector reliability is discussed as a limitation.